What it does
The state it sees is entirely semantic grades — health high/medium/low, distance far/mid/in range — with no numbers, because the documentation says Jev compares numbers poorly.
Apps & data pipelines
A 2D fighting demo where Jev calls the intent every few hundred milliseconds and frame-level code turns that intent into footwork, blocks and punches.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The state it sees is entirely semantic grades — health high/medium/low, distance far/mid/in range — with no numbers, because the documentation says Jev compares numbers poorly.
Start by grouping your game's health and distance into clear words instead of exact numbers. You can describe remaining health as high, medium, or low, and distance as far, mid, or in range. Sorting these details into everyday words helps the AI understand the match easily.
A developer can then connect your game to the online AI service. The model picks broad goals a few times each second, while your local code handles fast punches and blocks. If the connection fails or takes too long, the fighter simply stands still without an offline substitute.